스레드 생성 및 삭제 오버헤드를 줄이기 위한 히스토리 기반의 동적 스레드풀 방법

Abstract

In an environment with frequent job requests and short job processing times, thread pool methods are frequently used to increase throughput by reducing overheads due to thread creation and removal. A watermark method normally reduces unnecessary uses of resources by keeping the number of threads less than those needed in the maximum. In the absence of available threads, however, it processes jobs by creating additional threads up to a specified limit so that the system overhead increases due to creation of threads, which results in throughput degradation. This paper presents a history-based dynamic method that alleviates throughput degradation. By estimating and maintaining the number of threads needed for jobs, it reduces overheads due to thread creation and removal. According to experiments, compared to the watermark thread pool method, it shows average 33% increase in the number of threads kept and average 62% reduction in the number of threads created, which results in 6% increase in terms of system throughput.